KR20210059659A - Distance Estimation Method and Apparatus of UWB System Using Composite Neural Network - Google Patents

Abstract

The present invention relates to a distance estimation apparatus and method of an UWB system using a convolution neural network. The distance estimation apparatus includes a receiving part for receiving a transmission signal of a transmitter; a preprocessing part for converting the received signal into an image in a form of a two-dimensional matrix for training a convolution neural network based on an UWB channel model; and a learning part for performing learning with a convolutional layer on the image of the preprocessing part and outputs a result of learning as a distance estimate value between a transmitter and a receiver. By providing a distance estimation method based on a convolution neural network of an UWB system rather than a conventional distance estimation method based on a threshold, the distance estimation performance is superior to the distance estimation performance of the conventional technique.

Description

Distance Estimation Method and Apparatus of UWB System Using Composite Neural Network

The present invention relates to an apparatus and method for estimating a distance of an ultra-wideband system using a convolutional neural network, and more particularly, to a convolutional neural network (CNN) without estimating Time of arrival (TOA) using a conventional threshold. The present invention relates to a technique for estimating the distance of an ultra-wideband system.

Recently, various types of indoor positioning technologies have been used in various fields in various fields, and research is being actively conducted in companies and research institutes. Among indoor positioning technologies, an ultra-wideband-based system has a wide bandwidth of 500MHz or more in the frequency domain and a short pulse waveform in the time domain. Due to this feature, there is an advantage in that there is less interference of multi-paths and high resolution in the time domain has higher accuracy than other positioning techniques.

1 is a block diagram of a threshold-based distance estimation method widely used as a distance estimation method using a conventional ultra-wideband signal.

를 송신기가 전송한 초광대역 신호라고 가정하면, 수신기에서의 수신 신호

는 [수학식 1]과 같다.In Figure 1

Assuming that is the ultra-wideband signal transmitted by the transmitter, the received signal at the receiver

Is the same as [Equation 1].

[Equation 1]

는 채널 임펄스 응답이고, *는 컨볼루션이며,

는 잡음신호이고,

는 송수신기 거리에 따른 전파 지연시간을 의미하며, 본 발명의 일 실시예에서는 TOA라고 정의한다.here,

Is the channel impulse response, * is the convolution,

Is the noise signal,

Denotes a propagation delay time according to a transceiver distance, and is defined as TOA in an embodiment of the present invention.

The received signal is converted into a digital signal, which is shown in [Equation 2] below.

[Equation 2]

는 샘플링 주기이며,

는 샘플링 주파수이고,

은 수신신호의 관찰 길이이다. 이때,

은 추정하고자 하는 송수신기 사이의 최대 거리로 결정된다.here,

Is the sampling period,

Is the sampling frequency,

Is the observation length of the received signal. At this time,

Is determined as the maximum distance between the transceivers to be estimated.

에서 첫 번째 도착하는 신호를 검출하기 위해서는 문턱값을 결정해야 하는데, 최적의 문턱값은 SNR의 함수로 결정되며 SNR에 반비례한다. 따라서, 문턱값을 결정하기 전에 먼저 수신 신호의 SNR 추정이 선행되어야 한다. 이때, SNR은 다음의 [수학식 3]으로 표현된다.Receive signal

In order to detect the first signal that arrives at, a threshold value must be determined, and the optimal threshold value is determined as a function of SNR and is inversely proportional to SNR. Therefore, before determining the threshold value, the SNR of the received signal must be estimated first. At this time, the SNR is expressed by the following [Equation 3].

[Equation 3]

는 수신된 초광대역 신호의 평균 전력이며,

는 잡음의 평균 전력이다. 본 발명의 일 실시예에서는 평균 전력을 추정하는데 어려움이 있어

대신

의 최대값을 사용해 [수학식 4]와 같이 표현하였다.here,

Is the average power of the received ultra-wideband signal,

Is the average power of the noise. In an embodiment of the present invention, it is difficult to estimate the average power.

instead

It was expressed as [Equation 4] using the maximum value of.

[Equation 4]

에서 초광대역 신호가 존재하지 않는 시간 구간에서 잡음의 전력을 추정한다. 초광대역 신호가 존재하지 않는 시간의 첫 시점을

, 시간 길이를

이라 하면 SNR의 추정은 다음의 [수학식 5]와 같이 표현된다.And, the received signal

Estimate the power of the noise in the time interval in which the ultra-wideband signal does not exist. The first point of time when there is no ultra-wideband signal

, Time length

In this case, the estimation of the SNR is expressed as the following [Equation 5].

[Equation 5] (Lq -> Ln)

In addition, a threshold value is determined based on the estimated SNR as shown in [Equation 6] below.

[Equation 6]

는 양의 실수이며 채널 환경에 따라 최적화를 한다. 그리고

는 문턱값을 나타낸다. 이처럼 문턱값이 결정되면 TOA에 해당하는 이산 시간은 다음의 [수학식 7]과 같이 추정할 수 있고, 이에 따라 TOA는 다음의 [수학식 8]과 같이 도출할 수 있다.here,

Is a positive real number and is optimized according to the channel environment. And

Represents the threshold value. When the threshold is determined in this way, the discrete time corresponding to the TOA can be estimated as shown in [Equation 7] below, and accordingly, the TOA can be derived as shown in [Equation 8] below.

[Equation 7]

[Equation 8]

라고 하면 송신기와 수신기 사이의 거리는 다음의 [수학식 9]와 같이 도출된다.Also, the speed of light

If so, the distance between the transmitter and the receiver is derived as shown in [Equation 9] below.

[Equation 9]

However, the conventional distance estimation method has a disadvantage in that the distance estimation performance is low when the SNR is low or the distance of the transceiver is long. Accordingly, the present applicant intends to propose a method for estimating the distance of an ultra-wideband system using a convolutional neural network that has better performance than the conventional method.

Korean Patent Registration No. 10-1903218 (2018.09.20)

It is an object of the present invention to provide a convolutional neural network-based distance estimation method for indoor location positioning of an ultra-wideband system, which is superior to conventional techniques, rather than a conventional threshold-based distance estimation.

The distance estimation apparatus of an ultra-wideband system using the convolutional neural network of the present invention for achieving this technical problem,

A receiver for receiving a transmission signal from the transmitter;

A preprocessor for converting the received signal into an image in the form of a two-dimensional matrix for learning a convolutional neural network based on an ultra-wideband channel model; And

And a learning unit that performs learning by using the image of the preprocessor as an input image of the convolutional layer and outputs the estimated distance between the transmitter and the receiver.

A method for estimating a distance of an ultra-wideband system using a convolutional neural network according to an aspect of another embodiment of the present invention,

(A) the receiving unit generates a received signal for learning of the convolutional neural network; (B) converting the received signal into an image by a preprocessor; And (c) in which the learning unit learns the image received by the convolutional neural network, and the convolutional neural network derives an estimate of the distance between the transmitter and the receiver.

Preferably, after step (a),

Converting the received signal into a digital form;

Applying normalization to the received signal in the digital form; And

It may include converting the normalized one-dimensional received signal into a two-dimensional image form.

Preferably, step (c) is

A learning step of performing learning with a predetermined number of convolutional layers; And

It may include an estimating step of outputting distance estimates of one transmitter and a receiver as a pre-combination layer of a regression network with respect to the output of the convolutional layer.

Preferably in one convolutional layer of the learning step,

Using the convolution filter and the activation function Rectified Linear Unit (ReLU) for the received image,

It may be provided to perform max pooling on the operation result to reduce the size of data according to the operation result and sequentially transfer it to the next convolutional layer.

According to the present invention as described above, by providing a convolutional neural network-based distance estimation method for indoor location positioning of an ultra-wideband system, not the conventional threshold-based distance estimation, distance estimation performance compared to the conventional technique. Has an excellent effect on

1 is a block diagram showing a conventional distance estimation technique using a threshold value.
2 is a block diagram of a distance estimation technique to which a distance estimation method of an ultra-wideband system using a convolutional neural network is applied according to an embodiment of the present invention.
3 is an exemplary diagram illustrating a signal to which normalization is applied according to a distance estimation method of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention.
FIG. 4 is an exemplary diagram illustrating transforming a 1D vector signal into a 2D matrix after normalization according to a distance estimation method of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention.
5 is a diagram illustrating a structure of a deep learning model of a distance estimation method of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention.
6 is a diagram illustrating distance estimation performance according to SNR of a distance estimation method of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention.
7 is a diagram illustrating RMSE performance with respect to distance of a distance estimation method of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention.
8 is a flowchart illustrating a method of estimating a distance of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention.

Specific features and advantages of the present invention will become more apparent from the following detailed description based on the accompanying drawings. Prior to this, terms or words used in the present specification and claims are based on the principle that the inventor can appropriately define the concept of the term in order to describe his or her invention in the best way. It should be interpreted as a corresponding meaning and concept. In addition, when it is determined that a detailed description of known functions and configurations thereof related to the present invention may unnecessarily obscure the subject matter of the present invention, it should be noted that the detailed description thereof has been omitted.

Hereinafter, a method for estimating a distance of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention is as follows.

2 is a block diagram of an apparatus for estimating a distance of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention. Referring to FIG. 2, the apparatus for estimating a distance of an ultra-wideband system is a received signal for a transmission signal of a transmitter. And converting the received signal in a digital form into a digital form, transforming the converted digital form into an image form through normalization and shaping, and learning a received vector in the transformed image form to derive a distance estimate, Accordingly, the distance estimation apparatus may include at least one of the receiving unit 100, the preprocessing unit 200, and the learning unit 300.

, 채널 임펄스 응답을

라고 상정하면 수신부(100)의 수신 신호는 [수학식 1]과 동일하고 수신 신호는 전처리부(200)에 의거 디지털 형태로 변환되며, 디지털 형태의 수신 신호는 [수학식 2]와 동일하다. 이때 수신 신호는 절대값의 제곱이 취해지며 최대 크기가 1이 되도록 [수학식 10]을 통해 정규화(Normalization)를 수행하게 된다.In Fig. 2, the transmission signal

, Channel impulse response

Assuming that, the received signal of the receiving unit 100 is the same as in [Equation 1], the received signal is converted into a digital form based on the preprocessor 200, and the received signal in the digital form is the same as in [Equation 2]. At this time, the square of the absolute value of the received signal is taken, and normalization is performed through [Equation 10] so that the maximum size is 1.

[Equation 10]

As such, an example of a received signal to which normalization is applied through [Equation 10] is shown in FIG. 3. 3 is an example when the distance between the transceivers is 2m and the SNR is 20dB, the received signal to which normalization is applied is transformed into a 2D matrix, and the transformed 2D matrix satisfies Equation 11.

[Equation 11]

단위로 잘라 행렬

의 열을 구성한다. 따라서 행렬

의 크기는

가 된다.At this time, the length of the one-dimensional signal is

Truncated matrix

Make up the columns of. Hence the matrix

The size of

Becomes.

이고

이라면 50

40의 행렬로 변환되고, 도 4는 2차원 변환의 예를 보인다.For example

ego

If it is 50

It is transformed into a matrix of 40, and FIG. 4 shows an example of a two-dimensional transformation.

4B is a two-dimensional matrix representing a converted signal as a black-and-white image. The converted signal is input to a convolutional neural network, and the output is the distance between the transceivers.

5 is a structural diagram illustrating a convolutional neural network of the learning unit 300 according to an embodiment of the present invention.

As shown in FIG. 5, the size of the image of the preprocessor 200 is 50-by-40, and includes a total of 4 convolutional layers and 3 pooling layers.

In addition, after each convolutional layer, a batch normalization layer is formed, and in this case, the activation function uses a Rectified Linear Unit (ReLU). And, a model was constructed using one pre-combination layer and a regression layer.

Referring to FIG. 5, data 50x40 in the form of a preprocessed 2D image is transferred to the first layer through 8 channels for learning.

In addition, the size of the output data of the first layer including 3x3 filter size, 1 horizontal and 1 vertical stride, ReLU for activation function, and 72 weights is reduced based on Max pooling.

At this time, the size of Max pooling is 2x2, the stride is 2 in width and 2 in height, and the output data of Max pooling is transmitted to the second convolutional layer.

The second convolutional layer is composed of 16 channels, the filter size is 3x3, the stride is 1 horizontal and 1 vertical, the activation function is ReLU, and the number of weights is 1,152. Accordingly, the output data of Max pooling is learned. The output data of the second convolutional layer is reduced in size according to the second max pooling.

At this time, the Max pooling size is 2x2, the stride is 2 horizontal and 2 vertical, and the output data of Max pooling is input to the third convolutional layer.

And, the third convolutional layer consists of 32 channels, the filter size is 3x3, the stride is 1 horizontal and 1 vertical, the activation function is ReLU, and the number of weights is 4,608. Accordingly, the max pooling of the second convolutional layer is The output data is learned, and the output data of the third convolutional layer, which is a learning result, is reduced in size according to Max pooling.

At this time, the Max pooling size is 2x2, the stride is 2 horizontally and 2 vertically, and the output data of Max pooling is input to the fourth convolutional layer.

Meanwhile, the fourth convolutional layer consists of 64 channels, the filter size is 3x3, the stride is 1 horizontal and 1 vertical, the activation function is ReLU, and the number of weights is 18,432. The output data is learned, and the output of the fourth composite layer of the multiplication learning result is summed with the output of the fifth convolutional layer and input to the fully-connected layer. Here, the output of the pre-coupling layer is one because it is a regression network, and the output data is an estimate of the distance of the transceiver and is a real number.

Meanwhile, the distance estimation apparatus of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention may verify a distance estimation value derived by the learning unit 300.

과 실측값

간의 용함수의 최적해로 도출되며, 비용함수는 추정 거리값

과 실측값

에 대한 평균자승에러(RMSE : Root Mean Square Error)로 설정되며, 하기 수학식 12로 나타낼 수 있다.Here, the performance evaluation for the distance estimate is the estimated distance value.

And the measured value

It is derived as the optimal solution of the solution function between, and the cost function is the estimated distance value.

And the measured value

It is set as a root mean square error (RMSE) for, and can be expressed by Equation 12 below.

[Equation 12]

Here, B is the batch size of the convolutional neural network, and the optimal solution of this cost function is derived using the stochastic gradient descent method.

Hereinafter, a simulation result for confirming the performance of a distance estimation method of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention will be described with reference to FIGS. 6 and 7.

The simulation test environment is as follows. As the training data used in the deep learning model, a signal received through an ultra-wideband signal passing through the IEEE 802.15.4a channel is used. It is assumed that the sampling frequency of the received signal is 24 GHz and the carrier frequency is 3.5 GHz.

은 추정 최대 거리를 고려하여 25m에 해당하는

으로 설정하였고, 수신 신호 생성시 송신기와 수신기 사이의 거리와 SNR은 매번 설정된 범위 안에서 랜덤하게 결정된다.Observation length for the received signal

Is equivalent to 25 m, taking into account the estimated maximum distance

When the received signal is generated, the distance between the transmitter and the receiver and the SNR are randomly determined within the set range each time.

In addition, a total of 100,000 sets of received signals were generated and used as training data, and the SNR range was 10~30[dB], and the distance was randomly generated from 1~20m, the initial learning rate was 0.001, and the reduction in the learning rate was observed every time. Each interval (epoch) decreases by 0.9 times.

The maximum observation interval (epoch) is 60 and the mini-batch size is 200. The label is the transceiver distance, and learning is performed so that the output of the convolutional neural network becomes the transceiver distance.

Thereafter, when learning is complete, test data is generated to perform performance evaluation. Distance estimation performance evaluation according to SNR was performed in the same environment as the training data.

Performance evaluation was performed through RMSE comparison, and the RMSE in SNR is defined as shown in [Equation 12] below.

[Equation 12]

은 실제 송수신기 사이의 거리이고,

는 합성곱 신경망을 통해 추정한 거리이다. 그리고

은 데이터의 개수이다.here,

Is the distance between the actual transceivers,

Is the distance estimated through the convolutional neural network. And

Is the number of data.

6 is a diagram illustrating distance estimation performance according to SNR of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention.

Referring to the drawings, it can be seen that it shows superior performance than the conventional distance estimation technique, and in particular, shows an RMSE of 2m or less in all SNR ranges. In other words, it shows excellent RMSE performance regardless of SNR. On the other hand, it can be confirmed that the conventional distance estimation technique operates much more sensitively to SNR, and if a sufficiently high SNR is not secured, RMSE performance deteriorates a lot.

And, referring to FIG. 7, a similar RMSE performance can be confirmed regardless of the distance between the transmitter and the receiver, but according to the conventional method, it can be seen that the RMSE performance deteriorates as the distance increases.

Hereinafter, a method for estimating a distance of an ultra-wideband system using a convolutional neural network according to an embodiment of the present invention based on the background and configuration described above with reference to FIG. 8 will be described below.

First, the receiver 100 generates a received signal for learning a convolutional neural network based on an ultra-wideband channel model (S802).

Subsequently, the preprocessor 200 converts the received signal into an image in the form of a two-dimensional matrix for learning a convolutional neural network based on an ultra-wideband channel model, and converts the received signal into a digital form and then converts it into a digital form. The one-dimensional received signal of is converted into a two-dimensional image in the form of a matrix (S804). At this time, the size of the converted image is 50-by-40.

Subsequently, the learning unit 300 performs learning on the image received by the convolutional neural network (S806), and the convolutional neural network derives an estimated distance between the transmitter and the receiver as a result of the learning (S808).

According to an embodiment of the present invention, by providing a distance estimation method based on a convolutional neural network of an ultra-wideband system instead of distance estimation based on a conventional threshold, the distance estimation performance is superior to that of the conventional technique. have.

Although described and illustrated in connection with a preferred embodiment for illustrating the technical idea of the present invention as described above, the present invention is not limited to the configuration and operation as illustrated and described as described above, and deviates from the scope of the technical idea. It will be well understood by those skilled in the art that many changes and modifications are possible to the present invention without. Accordingly, all such appropriate changes and modifications and equivalents should be considered to be within the scope of the present invention.

Claims (7)

In the distance estimation apparatus of an ultra-wideband system using a convolutional neural network,
A receiver for receiving a transmission signal from the transmitter;
A preprocessor for converting the received signal into a matrix-type 2D image for learning a convolutional neural network based on an ultra-wideband channel model; And
And a learning unit that performs learning with a convolutional layer on the image of the preprocessor and outputs a distance estimation value between the transmitter and the receiver. The method of claim 1, wherein the pre-processing unit,
Convert the received signal into digital form,
It normalizes the received signal in digital form and converts it into one dimension,
An apparatus for estimating a distance of an ultra-wideband system using a convolutional neural network, characterized in that it is provided to convert the converted one-dimensional received signal into a two-dimensional matrix image. In the distance estimation method of an ultra-wideband system using a convolutional neural network,
(A) the receiving unit generates a received signal for learning of the convolutional neural network;
(B) converting, by a preprocessor, the received signal into a matrix-shaped 2D image for learning a convolutional neural network based on an ultra-wideband channel model; And
Distance estimation of an ultra-wideband system using a convolutional neural network, characterized in that the learning unit includes the step (c) in which the convolutional neural network derives an estimate of the distance between the transmitter and the receiver by performing learning on the image received by the convolutional neural network. Way. The method of claim 3, wherein after step (a),
Converting the received signal into a digital form;
Applying normalization to the received signal in the digital form; And
And converting the normalized one-dimensional received signal into a two-dimensional image form. The method of claim 4, wherein step (c)
A learning step of performing learning with a predetermined number of convolutional layers; And
And an estimating step of outputting a distance estimate of one transmitter and a receiver to a fully-combined layer of a regression network with respect to the output of the convolutional layer. The method of claim 5, wherein in one convolutional layer of the learning step,
A learning operation is performed with an activation function of a Rectified Linear Unit (ReLU) and a predetermined number of pooling coefficients for the received image,
A method for estimating a distance of an ultra-wideband system using a convolutional neural network, comprising performing maximum polling on an operation result, reducing the size of data according to the operation result, and sequentially transmitting it to the next convolutional layer. The recording medium according to any one of claims 3 to 6, wherein a program for executing a distance estimation method of an ultra-wideband system using a convolutional neural network is recorded and executable by a computer.

Images